GraphChef: Decision-Tree Recipes to Explain Graph Neural Networks
Authors: Peter Müller, Lukas Faber, Karolis Martinkus, Roger Wattenhofer
ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments demonstrate that Graph Chef reaches comparable accuracy to non-self-explainable GNNs, and the produced decision trees are indeed small. We further validate the correctness of the discovered recipes on datasets where explanation ground truth is available: Reddit-Binary, MUTAG, BA-2Motifs, BA-Shapes, Tree-Cycle, and Tree-Grid. |
| Researcher Affiliation | Academia | Peter Müller, Lukas Faber, Karolis Martinkus, Roger Wattenhofer ETH Zurich, Switzerland {lfaber,mkarolis,wattenhofer}@ethz.ch |
| Pseudocode | No | The paper describes the Graph Chef model and its components using textual descriptions, equations, and diagrams (e.g., Figure 2), but it does not include any explicitly labeled 'Pseudocode' or 'Algorithm' blocks. |
| Open Source Code | Yes | We provide a user interface for Graph Chef.1 This tool allows for the interactive exploration of the Graph Chef recipes on the datasets examined in this paper. We provide a manual for the interface in Appendix F. 1https://interpretable-gnn.netlify.app/ The supplementary material also contains code to host the interface yourself, in case you want to try variations of Graph Chef. |
| Open Datasets | Yes | We test Graph Chef on established GNN explanation benchmarks and real-world graph datasets. We use the Infection and Negative Evidence benchmarks from Faber et al. (2021), The BA-Shapes, Tree-Cycle, and Tree-Grid benchmarks from Ying et al. (2019), and the BA-2Motifs dataset from Luo et al. (2020). Second, we experiment with the following real-world datasets: MUTAG (Debnath et al., 1991); BBBP (Wu et al., 2017b); Mutagenicity (Kazius et al., 2005); PROTEINS, REDDIT-BINARY, IMDB-BINARY, and COLLAB (Borgwardt et al., 2005). |
| Dataset Splits | Yes | We do a 10 fold cross-validation of the data with different splits and train both Graph Chef and a baseline GIN architecture. GNNs are trained on the training set for 1500 epochs, allowing early stopping on the validation loss with patience of 100. Each split uses early stopping on the validation score. [...] We also further divide the training set for Graph Chef to keep a holdout set for pruning decision trees. |
| Hardware Specification | No | Note that all datasets except COLLAB are small enough to train on commodity CPUs. For example, training the PROTEINS dataset for one seed on a laptop trains in 5 minutes for the full 1500 epochs, a few seconds for the tree distillation, and 1–2 minutes for tree pruning. |
| Software Dependencies | No | In the backend, we use Py Torch (Paszke et al., 2019)3 and Py Torch Geometric (Fey and Lenssen, 2019)4 to train Graph Chef and SKLearn(Pedregosa et al., 2011)5 to train the decision trees. While specific libraries are mentioned, their exact version numbers are not provided, only the publication years of their respective papers. |
| Experiment Setup | Yes | Both GNNs use a 2 layer MLP for the update function, with batch normalization (Ioffe and Szegedy, 2015) and Re Lu (Nair and Hinton, 2010) in between the two linear layers. We use 5 layers of graph convolution. GIN uses a hidden dimension of 16, Graph Chef uses a state space of 10. We also further divide the training set for Graph Chef to keep a holdout set for pruning decision trees. After we train dish GNN with gradient descent, we distill the MLPs into decision trees. Each tree is limited to having a maximum of 100 nodes. |